A Spatio-Temporal Population Distribution Method for Emergency Evacuation:

A Case Study of New Orleans

Ruijie Bian*

Department of Civil and Environmental Engineering

Louisiana State University

2400 Patrick F. Taylor Hall, Baton Rouge, LA 70803

Email: rbian1@tigers.lsu.edu

Chester G. Wilmot

Department of Civil and Environmental Engineering

Louisiana State University

3513C Patrick F. Taylor Hall, Baton Rouge, LA 70803

Email: cecgw@lsu.edu

(*) Corresponding Author

Submission: August 1, 2014

Word count: 5450 words+ 2 Tables + 6 Figures (250 words per table/figure) = 7450 total words

ABSTRACT

Knowing where people are at different times of the day is important in emergency management.

This study presents a spatio-temporal method to estimate population distribution by time of day,

day of week, and season of the year. Population is broken into six groups: resident, worker,

student, stay-at-homer, shopper, and tourist. The last two groups have usually been neglected in

past studies. However, they can be significant portions of the population in certain environments

and display unique dynamic daily/weekly migration and seasonal variation. In this study,

population is distributed using dasymetric mapping where distribution of the population is based

on land use in the area. However, two modifications are made to the usual manner in which the

process is applied. First, preprocessing of coarse land use categories with residential density and

employee data allow distinction between resident, worker, shopper, and tourist in the distribution

process. Second, the relationship between population and land use is based on regression of

existing data instead of on subjective judgment or sampling as in current dasymetric methods.

The method is demonstrated in an application to downtown New Orleans evaluating the

population affected by a hypothetical chemical spill. In one case, the spill is assumed to occur on

a weekday afternoon during a local festival. Over 82,813 people are estimated to be affected by

the toxic plume emanating from the site, with most being workers in the industrial area

surrounding the port as well as tourists in the French Quarter. In the other case, the same spill

was assumed but considered to occur over the weekend at night during a period when no festival

was in progress. In this case, only 39,147 people were estimated to be affected with most of them

being residents and tourists spread throughout the area within the plume.

Bian and Wilmot 1

INTRODUCTION 1 Natural (e.g. hurricane, flood, and earthquake) and man-made (e.g. chemical spill) disasters can 2

have huge economic and humanitarian costs. For example, hurricane Katrina caused damage 3

estimated at over $100 billion, killed approximately 1,800 people, and displaced over 1 million 4

from their homes [1] [2]. The gas leak in Bhopal, India, in 1984 resulted in the death of over 5

8,000 [3]. Of course, major disasters such as these are rare, but when all disasters are considered 6

collectively at national or global level, the importance of disasters to society as a whole is 7

obvious. For example, during the past 10 years (2004 - 2013), 1,367 disasters have occurred in 8

the United States alone [4]. Among these, weather-related disasters have accounted for 6,403 9

fatalities [5]. 10

One way to counter the harm caused by a disaster is to move people away from its impact. When 11

disasters have long warning times, such as with an approaching hurricane, most evacuation takes 12

place from home as people have time to coordinate their individual activities into a household 13

plan. Thus, a good estimate of the location and magnitude of demand for evacuation can be 14

obtained from sources such as the Census which provides population estimates at the home. 15

However, no-notice and short-notice disasters can occur when people are away from their homes 16

and the population distributions are then potentially quite different from the residential 17

population distribution. Estimating the spatio-temporal location of people by time of day, day of 18

week, and season of the year in a metropolitan area, is the subject of this paper. 19

Some studies have observed daily migratory patterns by tracking the path and activities of people 20

from their smartphone data [6]. However, in 2013 smartphone users in the U.S. made up only 21

about 56 percent of American adults [7]. The ownership possibility for elderly (more than 65) is 22

18% and for low-income (less than $30,000/year) is 43% [7]. The possibility of owning a 23

smartphone for disabled people is even lower. At the same time, these people are those most 24

likely to need help during an evacuation. 25

Estimating the movement of all individuals in a metropolitan area would be a daunting task, but 26

groups of people with similar characteristics tend to have similar movement patterns and time 27

schedules, which makes estimation of the temporal distribution of people more feasible. In past 28

studies, the population has been divided into groups such as residents, workers, and students [8] 29

[9]. However in cities of historic, cultural, or entertainment interest, tourists are another 30

important population group. Tourists vary across seasons and peak during holidays and festivals. 31

People who do not work or go to school remain at home or participate in other activities such as 32

shopping, visiting friends, or participating in recreational activities. Thus, stay-at-homers are 33

another group of the population that deserves consideration but they, together with tourists and 34

shoppers, have usually been neglected in past studies. The spatio-temporal population 35

distribution of all of these groups are likely to be different between a weekday and a weekend, so 36

variation by day of week also deserves attention. 37

Intuitively, knowledge of the land use of an area can provide a better estimate of the population 38

distribution since land use dictates activity type. However, uniform and detailed land use maps 39

are hard to find for an entire metropolitan area. Detailed land cover/land use maps are usually 40

compiled at local level, so areas differ on the level of detail and may use different land use 41

categories. The most commonly used land use categories are those used in the National Land 42

Cover Database (NLCD). However, the shortcoming of the categorization used in the NLCD is 43

that, while it distinguishes land use by intensity, it does not distinguish between residential, 44

Bian and Wilmot 2

industrial, and commercial land uses. The result is difficulty in assigning residents, workers, 1

tourists, and shoppers to areas where the intensity suggests there is much activity, but the nature 2

of the activity is unknown. The solution to this problem is to use ancillary data to improve the 3

distribution. 4

In this study, a spatio-temporal population distribution process, which accesses open source data 5

and modifies an existing dasymetric mapping method, is presented. The population distribution 6

at any moment is assumed to depend on the time of day, day of week, and season of the year. In 7

addition, to more fully cater for the diversity of behavior in the population, people are broken 8

down into six categories: resident, worker, student, stay-at-homer, shopper, and tourist. 9

The following section provides a literature review of past studies on spatio-temporal distribution 10

estimation, including how spatial and temporal aspects were handled and population groups 11

established. The next section presents the methodology used to distribute the population using 12

ArcGIS as the platform to conduct the procedure. Following this, a case study is presented in 13

which a hypothetical chemical spill in New Orleans is used to demonstrate the spatio-temporal 14

significance of no-notice or short-notice evacuation demand. Conclusions and future 15

improvements are discussed in the final section of the paper. 16

LITERATURE REVIEW 17 Past studies show great variety in treatment of the spatial, temporal, and population breakdown 18

of dynamic population estimation. For example, some studies focus on estimating a single 19

population group, such as tourists, on a monthly level [10]. Though the estimates are at state 20

level, it still required considerable time and effort using a survey to achieve the results. Kellens 21

et al. [11] also estimated tourist distribution for a coastal area, but on a seasonal level. 22

Other studies have considered smaller spatial areas and shorter temporal variation. This has 23

usually required user’s intuitive judgment in supplying certain data or relied on acquiring the 24

information from special surveys. For example, Zhang et al. [12] developed a spatio-temporal 25

model using building category as the spatial classification factor and divided the day into 17 time 26

periods. Though the spatial classification was relatively detailed, the population density of each 27

building category by time period was based on subjective assessment. To achieve more objective 28

results, Ahola et al. [13] conducted a survey to estimate population density as a function of 29

building category, time of day, and age of the respondent. The period of a week was divided into 30

14 periods. In each time period, the population of 10 age-based categories were expressed as a 31

percentage of the total population. However, most of these percentages were based on user 32

judgment because the survey was unable to provide reliable estimates for all categories. Bell [14] 33

used a parking space survey for daytime population estimation in each block. LandScan Global 34

[15], developed by Oak Ridge National Laboratory, uses a Geographic Information System 35

(GIS) and a Remote Sensing method to estimate global population distribution. The spatial 36

dimension used in LandScan Global is 1 km resolution, which is helpful at global level, but quite 37

coarse for a local area. 38

As regards the temporal dimension, most studies consider only a rough breakdown: nighttime 39

and daytime of a single day. Nighttime population is usually considered to be the resident 40

population. Studies vary in terms of daytime population breakdowns. ESRI offers a thematic 41

map of “USA Daytime Population” which only considers workers in an area. McPherson and 42

Brown [16], researchers at Los Alamos National Laboratory, developed daytime distributions 43

Bian and Wilmot 3

based on worker data and population remaining home during the daytime. They further improved 1

their study by introducing fractions of people indoors and outdoors [17]. Later, they added 2

population exchange elements into their study [18]. Population exchange elements are the 3

proportion of residents who live in one zone and work in another zone, and the proportion of 4

workers who work in one zone and live in another. Workers and students are typical subsets of 5

the population considered in daytime population distributions, such as in the studies of Freire [8] 6

[9], Freire and Aubrecht [19] [20] [21], and Sleeter and Wood [22]. 7

Regarding the spatial dimension, study areas vary from local to global levels, and sometimes 8

even involve multiple levels [23]. The method usually used to disaggregate the population is the 9

areal interpolation method. This involves disaggregating the population on area alone. A second 10

common method is pycnophylactic interpolation where a smooth mathematical function is used 11

to provide areal weighting to the areal interpolation method [24]. Kobayashi et al. [25] used this 12

method in their study. Another method is dasymetric mapping, or intelligent dasymetric 13

mapping, which disaggregates population in accordance with ancillary data such as length of 14

street centerline or square feet of certain land uses within the area. Freire [8] [9], Freire and 15

Aubrecht [19] [20], and LandScan use this method in their research. 16

This study uses dasymetric mapping method with some modifications. Dasymetric mapping is an 17

areal interpolation method that uses ancillary spatial data in its interpolation. Land cover/land use 18

maps are most often used as ancillary data. Holloway et al. [26] used a fixed proportion of the 19

population for each land use based on subjective judgment. Land use types used in their study 20

included urban, open, agriculture and forest, and uninhabited. Mennis [27] combined land use 21

sampling with areal-based weighting in their dasymetric mapping procedure. In that research, 22

land use types are based on urbanization level, which includes uninhabited, non-urban, low-23

density residential, and high-density residential. For each land use, population density values 24

were sampled using a survey. The population was then distributed based on a weighting factor 25

related to population density values and area percentage of each land use within one geographic 26

unit. Later, Sleeter and Gould [28] developed an ArcGIS-based Dasymetric-Mapping Extension 27

(DME) procedure based on the Mennis [27] theory. 28

Limited land use types and subjective judgment on the relative importance of each land use type 29

are two problems that limit the usefulness of the aforementioned methods. However, it is known 30

that land use is related to population so if land use distribution is known at a finer level of detail 31

than population, it can be used to infer population distribution. 32

Besides, in most past studies tourists and shoppers have been ignored as population subgroups 33

with distinct spatio-temporal patterns. Even though Bhaduri [24] and Bhaduri et al. [29] consider 34

tourists in the LandScan USA study, they do not consider the seasonal variation and the peaking 35

in attendance during festivals or other special events. To a large extent, shoppers have been 36

ignored in past studies even though national travel studies such as the National Household Travel 37

Survey (NHTS) show that shopping trips can make up to 25.7% percent of a household’s daily 38

travel. 39

In this study the population is broken into six groups – residents, workers, students, shoppers, 40

stay-at-homers, and tourists. On the temporal dimension, a day is divided into daytime and 41

nighttime, a week into weekday, Saturday, and Sunday, and a year into four quarters 42

corresponding to the four seasons of the year. Special consideration is also given to periods when 43

festivals or other special events occur in the area under consideration. For the spatial dimension, 44

Bian and Wilmot 4

the Traffic Analysis Zone (TAZ) polygon layer, commonly used by Metropolitan Planning 1

Organizations (MPOs), and the NLCD land use converted polygon layer are intersected to 2

establish a new set of spatial units to which population groups are assigned. 3

The NLCD land cover/land use map provides uniform land use categories across the whole 4

country. Its one disadvantage is that while it does distinguish land use by intensity, it does not 5

distinguish between residential, commercial, and industrial activity within each intensity 6

classification. To overcome this disadvantage, residential density and employee data are 7

collected from other sources to distinguish between the land uses of different population groups. 8

METHODOLOGY 9

Input data 10

The first task in developing a spatio-temporal estimate of the population of an area is to identify 11

the source and geographic level at which input information on the characteristics of the 12

population are available. That is, where can data on population, employment, students, tourists, 13

commerce, and industry be obtained and at what geographic level is it available? In this study, 14

data from the Census Transportation Planning Package (CTPP) was used to provide information 15

on population and employment at TAZ level. Data from the National Center for Education 16

Statistics (NCES) was used to provide student numbers by educational institution. The National 17

Household Travel Survey (NHTS) was used to estimate shopping frequency and the time of day 18

when shopping occurs, County Business Pattern (CBP) data of employment in the retail and 19

hospitality industry was used to improve the precision of distributing shoppers to the shop end of 20

their travel and the location of tourists at night. Information from Convention and Visitor 21

Bureaus (CVBs), as well as information extracted from local newspaper articles, were used to 22

estimate the number of tourists, their distribution by city and season, and their peaking at the 23

time of regularly scheduled festivals. TABLE 1 summarizes the sources and geographic level of 24

data used as input in this study. 25

TABLE 1 Spatial Data and Sources 26

Group Data/Ancillary data source Data level Data type

Household population CTPP-residence TAZ Polygon

Household employment CTPP-residence TAZ Polygon

Household school enrollment CTPP-residence TAZ Polygon

Worked at home CTPP-residence TAZ Polygon

Stay-at-homer (basic) CTPP, NCES1 TAZ Polygon

Worker at workplace CTPP-workplace TAZ Polygon

Student at school/university NCES Specific address Point

Shopper at home-end NHTS State Polygon

Shopper at shop-end CBP Zip code Polygon

Daytime tourist CVB, local newspaper Parish Polygon

Nighttime tourist CBP Zip code Polygon

Land cover/land use map NLCD Raster

1 Stay-at-homers (basic) are calculated according to the equation “Residents – Students –

Workers”.

Bian and Wilmot 5

Estimating Residents 1

Residents are assumed to be at home at night. During the day, some residents participate in out-2

of-home activities and therefore reduce the daytime residential population. For example, workers 3

and students are assumed to work and attend school during the day on weekdays. During the day 4

on weekdays, stay-at-homers are estimated by subtracting workers and students from residents 5

with allowance made for some stay-at-homers who go shopping during the day. During the day 6

over weekends, stay-at-homers are estimated by subtracting shoppers from residents. At night, 7

stay-at-homers are assumed to equal the number of residents. 8

Estimating Shoppers 9

The temporal distribution of the shopping activity is based on NHTS data which provides home-10

based shopping trip start and end times of person trips across the country. The person expansion 11

factor is also provided so the percentage of people who go shopping in one state or even over the 12

whole country during different times of the day or day of the week can be calculated. In this 13

study, a day is divided into daytime and nighttime periods. For daytime shopping, shopping was 14

assumed to occur between 10 am and 6 pm as evidenced by the time distribution of shopping in 15

the NHTS data in FIGURE 1. Shopping activities in this time period averaged about 66% of an 16

entire weekday’s shopping activities. From the NHTS data it was also determined that the 17

proportion of the population who go shopping on any particular day are 4.6% of the population 18

on weekdays, 0.9% on Saturdays, and 0.6% on Sundays. 19

20

FIGURE 1 Temporal Distribution of Shopping Trips 21

People are assumed to go shopping during the daytime only. Shoppers were estimated as a 22

proportion of stay-at-homers during the week and a proportion of the whole population during 23

the weekend. 24

The distribution of shoppers at the home end is based on the proportion of residential area in 25

each TAZ. For the shop end, if data from the NLCD land cover/land use maps are used in the 26

distribution process, shoppers could be disproportionally distributed to residentially concentrated 27

areas instead of main commercial areas as they both belong to the same high intensity land use 28

type in the NLCD classification. So, more ancillary data are needed to improve the assignment 29

0

10000000

20000000

30000000

40000000

50000000

60000000

70000000

80000000

Trips Weekday

Saturday

Sunday

Bian and Wilmot 6

process. In this study, based on a constant service/customer ratio assumption, it is assumed that 1

the number of shoppers is proportional to the number of retail employees. Retail employee data 2

are collected from CBP. At zip code level, the original data file only provides the number of 3

establishments and the range of number of employees. The midpoint of each range was used for 4

the number of employees in this study. However, since specific retail employee data is available 5

at parish level, balancing was done to match the estimated retail employees at zip code level to 6

the values at parish level. Following this, the total shoppers at the shop end were assumed to be 7

equal to the total shoppers at the home end. Then, shoppers were distributed proportionally to the 8

number of retail employees in each zip code area. By doing this, the precision of shopper 9

distribution was controlled at least at zip code level. Within a zip code area, the modified 10

dasymetric mapping method described below was used to distribute shopper to each kind of land 11

use. 12

Estimating Tourists 13

Tourist data is usually only available at state level with seasonal variation. Spatial ancillary data 14

such as the share of tourism going to major areas within the state can help distribute tourist to 15

each parish more precisely. Temporal ancillary data like festival news reports in local 16

newspapers can help in estimating high volumes for festival periods and low volumes for non-17

festival periods in each season. In this study, festival participants of a season are summed and 18

averaged by festival days of that quarter. As an average number is used in this study, actual 19

attendance during peak periods of major festivals may exceed this average number but in the 20

absence of knowing daily attendance at individual festivals, this is a practical solution. Tourist 21

numbers on non-festival days are estimated by subtracting total festival participants from the 22

total number of tourists of that quarter and averaged over the non-festival days. 23

In order to improve the precision of nighttime tourist distribution, hotel employee data are 24

collected from CBP. And similar to the distribution of shoppers to the shop end, nighttime tourist 25

were distributed proportionally to the number of hotel employees. So, its precision is also be 26

controlled at zip code level. 27

Population Distribution with the Modified Dasymetric Mapping Method 28

Distinguish Industrial Area 29

The most accessible and commonly used land cover/land use map is National Land Cover 30

Database 2011 (NLCD 2011). In this database, land use is classified into 20 land use types and 31

the data is provided at a resolution of 30 m×30 m. In each 30 m×30 m area only one land use is 32

recorded. The advantage of this database is that it provides a uniform land use classification 33

across the entire U.S. However, it does not provide a detailed division of urban land use. For 34

example, the high intensity land use types in NLCD include residential (e.g. high rise apartment 35

buildings), industrial, and commercial areas. For purposes of this study, industrial areas cause the 36

biggest problem in this classification as they do not serve residents or shoppers. So, industrial 37

areas must be separated from other high intensity land use. This was achieved in this study by 38

calculating the residential density of high intensity land use areas at the TAZ level, and noting 39

their densities. Densities of some known industrial areas were then reviewed to set up a threshold 40

below which areas are clearly industrial. TAZs with density values above the threshold were 41

considered residential or commercial. 42

Bian and Wilmot 7

Relative Importance of Land Use 1

Instead of using subjective judgment on the percentage of population to be assigned to each land 2

use type in dasymetric mapping, linear regression can be used to find the relative importance of 3

each land use. If population is taken as the dependent variable and the areas of the different land 4

use types are taken as independent variables, the regression coefficients will express the 5

population associated with each land use type per unit area. If the same unit of area is used in 6

each independent variable, the coefficients will express the relative weight of each land use type 7

in attracting population. This approach is used in the next step to distribute population within 8

each geographic unit (i.e. TAZ, Zip code, and Parish) to land use type within that geographic 9

unit. 10

Disaggregate Population 11

The simplest assumption regarding the distribution of population in an area is to assume they are 12

distributed uniformly throughout the geographic unit. However, some areas within the 13

geographic area may be uninhabitable and others may be less attractive for housing, so a uniform 14

assignment of population throughout the area would be inappropriate. The dasymetric mapping 15

method addresses this by distributing population by land use within each geographic unit. In this 16

study, the association between land use and population is quantified using data where population 17

and the area of different land use types is known by geographic unit. This is typically the case for 18

geographic units at TAZ, ZIP code, or parish level. Using this data, each population group at its 19

original level (e.g. resident at TAZ level) is regressed against the area of the different land use 20

types in each geographic unit. 21

𝑃𝑜𝑝𝐺𝑈𝐿𝑈𝑖

𝑝 = 𝑃𝑜𝑝𝐺𝑈𝑗

𝑝 ×𝑐𝑜𝑒𝑓𝑓𝐿𝑈𝑘

𝑝.𝑎𝑟𝑒𝑎𝑖,𝑘

∑ 𝑐𝑜𝑒𝑓𝑓𝐿𝑈𝑘𝑝

.𝑎𝑟𝑒𝑎𝑖,𝑘.𝛿20𝑘=1

(1) 22

𝑃𝑜𝑝𝐺𝑈𝐿𝑈𝑖

𝑝: Number of persons estimated to belong to population group p in intersection polygon 23

𝑖. The intersection polygon is the result of intersecting the geographic unit layer and the land use 24

layer; 25

𝑃𝑜𝑝𝐺𝑈𝑗: Number of persons belonging to population group p in geographic unit polygon 𝑗. 26

Intersection polygon 𝑖 is within geographic unit polygon 𝑗; 27

𝑐𝑜𝑒𝑓𝑓𝐿𝑈𝑘

𝑝: Parameters of land use type 𝑘 estimated in the regression equation for population 28

group p. Intersection polygon 𝑖 has land use attribute 𝑘; 29

𝑎𝑟𝑒𝑎𝑖,𝑘: Area of land use type k in intersection polygon i; 30

𝛿: 𝛿 equals 1 when land use type 𝑘 appears within geographic unit polygon 𝑗. Otherwise, it 31

equals zero. 32

Time Elements 33

Using equation (1) for each population group, the population in each geographic unit was 34

distributed to the intersection polygons. Intersection polygons fit within the geographic units and 35

consist of a single land use type. For each time period, related data should be summed together 36

from all population groups to get the total population distribution during that period. For 37

example, an estimate of the first quarter festival weekday population in an area would be the sum 38

of the following for all intersection polygons in the area: 39

Bian and Wilmot 8

Weekday nighttime population = nighttime resident (household population) + nighttime tourist 1

of the first quarter festival 2

Weekday daytime population = worker + student + weekday stay-at-homer + weekday shopper + 3

daytime tourist of the first quarter festival 4

The difference between a weekday and a day during the weekend is the daytime 5

population distribution. Take a Sunday in the first quarter as an example: 6

Sunday daytime population = Sunday stay-at-homer + Sunday shopper + daytime tourist of the 7

first quarter festival 8

In this study, for a whole year, 32 different time periods are finally calculated. For example, 9

there are 8 time periods in each quarter, in which 6 are in the daytime and 2 are at night. The 6 10

daytime periods are determined by whether the day is a festival day or not, and whether it is a 11

weekday, Saturday, or Sunday. The 2 nighttime periods are only distinguished by whether or not 12

it is a festival day. 13

CASE STUDY 14 New Orleans, Louisiana, was chosen as the site to demonstrate the procedure described above. 15

The population of the New Orleans metropolitan area is approximately 1.2 million but it attracts 16

a large number of tourists. The Louisiana Visitor Profile Report lists the number of tourists 17

visiting Louisiana in 2012 at over 26 million and New Orleans is estimated to host 18

approximately 38% of that number. Although the duration of their visit is not known, if it were 19

one day, that would result in an average of 27,000 tourists per day in the New Orleans 20

metropolitan area with a higher number during festivals and a lower number during off-peak 21

periods. Thus, tourists can make up a significant portion of the population in New Orleans during 22

particular times of the year. Knowing how many people are in New Orleans and where they are 23

at any time is an important issue for emergency preparedness and the evacuation process. 24

The Orleans parish land use map is shown in FIGURE 2. Note, Orleans parish is just one parish 25

within the six-parish New Orleans metropolitan area but is used here to more easily distinguish 26

the small intersection polygons and the estimated population in them. The map displays 15 types 27

of NLCD land use in the Orleans parish area. As seen by the blue in the map, open water and 28

wetlands occupy a large portion of the included area. If the population were to be distributed 29

uniformly within the area, people would be distributed to open water and wetlands. Thus, the 30

dasymetric approach which uses land use information in the distribution of the population is 31

more appropriate. However, the only land use classification in NLCD that is associated with 32

human occupation is “developed land” and it is only distinguished by intensity of human 33

presence (high, medium, and low intensity, and open space). This does not distinguish whether 34

people are living, working, shopping, or visiting in these areas. As mentioned earlier, this is a 35

problem in high density industrial areas since, if people are distributed by land use intensity, 36

residents, tourists, and shoppers could as easily be distributed to high intensity industrial areas 37

such as a port or industrial park as workers could be assigned to high rise residential areas. This 38

was addressed in this case study by observing the population density in TAZs that are known 39

industrial areas and comparing the population density with those that were not industrial areas. In 40

this study the cutoff distinguishing industrial from other land use was found to be 0.0041 persons 41

per square meter. This was used to distinguish industrial from other land use in the high density 42

Bian and Wilmot 9

category; no distinction was made in medium and low density cases. The land use area expressed 1

as a percentage for each land use category is shown in the second column of TABLE 2. 2

3 FIGURE 2 Land use of New Orleans. 4

Regression analysis was conducted on the data of the variables listed in the first column of 5

TABLE 2. The dependent variable is the number of persons in each population group in the 6

geographic unit at which the population is known (e.g. TAZ, ZIP code, or parish area). The 7

independent variables are the areas of different land use types from NLCD aggregated to TAZ 8

level. Regression results are listed in the last five columns of TABLE 2. All coefficients have the 9

expected sign and are significant at the 95 percent level of significance. 10

With the coefficients listed in TABLE 2, population can be distributed by land use. In order to 11

keep the final data at the same geographic unit level, TAZ and land use intersection polygons 12

(TAZLUs) are used as the basic geographic unit layer. Population data at TAZ level, such as 13

resident and worker, can be disaggregated to intersection polygon level by using equation (1). 14

However, for population data at other levels, such as nighttime tourist at zip code level, 15

additional steps in the process are necessary. First, intersect the TAZLU with Zip code layers. 16

Second, apply equation (1) to the TAZLU_ZIP intersection layer. Third, summarize the data to 17

TAZLU level. 18

Bian and Wilmot 10

1

TABLE 2 Land Use Type and Regression Result 2

Land Cover (ha) Area

(%)

Resident Worker Shopper Nighttime

Tourist

Daytime

Tourist

Coeff. t-

value Coeff.

t-

value Coeff.

t-

value Coeff.

t-value Coeff.

t-

value

Open Water 51.46 0 0 0 0 0

Developed, Open Space 0.84 2.9942 10.48 7.084 4.51 0 0 3.5971 3.24

Developed, Low Intensity 12.24 4.2378 18.57 0 0.0922 2.77 0.9911 2.81 0

Developed, Medium Intensity 5.7 18.3 14.23 0 0.5868 2.95 5.8165 2.77 0

Developed, High Intensity,

Industrial 2.05 0 10.5 15.41 0 0 0

Developed, High Intensity,

Others 0.89 51.2 15.95 55.9 36.4 2.8988 4.06 11.8 1.59 59.5 4.13

Barren Land 0.12 0 0 0 0 0

Deciduous Forest 0.03 0 0 0 0 0

Evergreen Forest 0 0 0 0 0 0

Mixed Forest 0 0 0 0 0 0

Shrub/Scrub 0.08 0 0 0 0 0

Grassland/Herbaceous 0.07 0 0 0 0 0

Hay/Pasture 0.25 0 0 0 0 0

Cultivated Crops 0.29 0 0 0 0 0

Woody Wetlands 5.75 0 0 0 0 0

Emergent Herbaceous Wetlands 20.15 0 0 0 0 0

No. of Observations 2124 2129 185 184 19

R square 0.6486 0.4406 0.5854 0.4338 0.7195

3

4

5

6

Bian and Wilmot 11

Population Group Distribution Analysis 1

Three dasymetric mapping methods distributing nighttime residents are compared to illustrate 2

the difference in results obtained with different methods. The results of distributing residents 3

with the Holloway method (which uses a fixed proportion of the population for each land use 4

based on subjective judgment) is shown in FIGURE 3 (a). The Dasymetric-Mapping Extension 5

(DME) method, which uses a sampling method to establish population weights for each land use 6

but assigns population at TAZ level, is shown in FIGURE 3 (b). Among these two methods, the 7

Holloway method clearly provides a more detailed estimate of population distribution but both 8

methods fail to avoid distributing nighttime population to industrial areas such as the port area 9

located on the western side of the crescent of the Mississippi river. FIGURE 3 (c) shows the 10

results of the method used in this study. It shows the greatest detail of population distribution and 11

residents are not distributed to the port. However, all methods solve the problem of distributing 12

population to open water area. 13

(a) Resident distribution by Holloway et al.

method [26]. Unit: Resident per cell.

(b) Resident distribution by DME [27].

Unit: Resident per cell

(c) Resident distribution by modified method of this study. Unit: Resident.

FIGURE 3 Resident Distribution by Different Method

14

Bian and Wilmot 12

Student data are collected at point level and therefore do not require spatial distribution. At the 1

temporal level they are assumed to be at school during the daytime on weekdays and at home the 2

rest of the time with the exception of the portion expected to participate in shopping over the 3

weekend. Worker distribution is shown here to illustrate the precision of the method used in this 4

study. As FIGURE 4 shows, workers are concentrated in the port and downtown commercial 5

area of the city. 6

7 FIGURE 4 Worker distribution 8

In this study, “high intensity other” land use includes high density residential, 9

commercial, and mixed land use areas. To distribute nighttime tourists and shoppers directly to 10

this land use may cause a problem due to the diversity of the land uses it represents. Employee 11

data are used as ancillary data in this study to improve the appropriateness of the distribution. 12

Specifically, the retail employment is used to identify commercial areas, and hotel employee data 13

is used to identify the spatial location of hotels. FIGURE 5 (a) shows shoppers distributed to 14

retail-employee concentrated areas. FIGURE 5 (b) shows nighttime tourists distributed 15

according to the distribution of hotel employee concentration in the area. 16

Daytime tourist distribution is based on analysis of local conditions. In New Orleans, 17

festivals occur mainly in Open Space and High Intensity Other land use areas, such as parks and 18

mixed commercial areas. Thus, daytime tourists are distributed in relation to these land uses as 19

shown in FIGURE 5 (c). 20

21

Bian and Wilmot 13

(a) Weekday shopper distribution (b) Nighttime tourist distribution

(c) Daytime tourist distribution

FIGURE 5 Shopper and Nighttime Tourist Distribution

Chemical Spill Case 1

To illustrate how the method developed in this study can be used, the case of a hypothetical 2

chemical spill in New Orleans is considered. The affected area is developed based on guidelines 3

in the Emergency Response Guidebook 2012 [30]. The location of the spill, chemical material 4

type, wind bearing, and scale are the same in two cases. The location of the spill and the toxic 5

plume that developed is shown in FIGURE 6. Two different times of occurrence are assumed to 6

illustrate the impact of time of day on who is affected. Say in the first case (case 1), the spill 7

occurred on March 5, 2014, at 2 pm. This is a weekday in the first quarter of the year during a 8

festival period, and it's a daytime occurrence. How many people will be exposed to the toxic 9

gas? In case 2, assume the same chemical spill occurred almost a month earlier on a Saturday 10

night in a non-festival period. Note that while this occurrence is still in the first quarter of the 11

year, it is on a weekend and at night. How many people would be affected in this case? 12

Case 1: Daytime population is the sum of worker, student, weekday stay-at-homer, 13

weekday shopper, and daytime tourist of the first quarter during a festival period. 14

School/university data are specific to location, which present as points in ArcGIS. It is easier to 15

judge how many students are affected by summing the point layer. And whether it is during a 16

Bian and Wilmot 14

school season when the event happen must also be taken into account. Assume when this event 1

occurs, schools are closed. So, population groups except students are summed. See FIGURE 6 2

(a). 3

Case 2: Nighttime population is the sum of residents and nighttime tourists of the first 4

quarter non-festival days. See FIGURE 6 (b). 5

Applying the procedure developed in this study, for case 1 the GIS system is able to sum the 6

population within the plume area for the specified conditions. It was found that the affected 7

number of population for Case 1 was 82,813. Looking at the distribution of the affected 8

population in FIGURE 6 (a), it is clear they are concentrated in the downtown commercial and 9

industrial areas. The number of workers was 58,335, stay-at-homers was 8,490, shoppers was 10

1,001, and tourists was 14,988. Thus, the majority of people affected are workers at their place of 11

employment. For case 2, the results are shown in FIGURE 6 (b). The total affected number of 12

population is estimated at 39,147, and the population is much more dispersed than in case 1. The 13

estimated number of residents was 30,061 and tourists was 9,086. Note that the largest group 14

affected is now residents. Clearly, the number and composition of those affected varies 15

considerably with time and the method is able to estimate the changing conditions over time. 16

(a) Case 1 (b) Case 2

FIGURE 6 Population Distribution in Two Cases

Information of this nature can be useful to emergency managers during a chemical spill or other 17

kinds of similar emergencies. Affected population distribution estimation can help organize 18

effective evacuation and rescue, such as on-spot rescue resources and hospital resources 19

preparation. 20

CONCLUSION 21 The objective of this study is knowing where people are during different times of the day, days 22

of week, and seasons of the year. Activities during the day, such as work and education, result in 23

movement from residential places to other areas. This has been studied thoroughly in past 24

studies. During the weekend, work and education needs decrease, but shopping and other 25

activities increase. This has an impact on stay-at-homer numbers and underscores the importance 26

of including all significant population groups such as shoppers and tourists in cities where 27

tourism is an important part of the cities activities. 28

Bian and Wilmot 15

How many people there are in each population group and how to distribute them was the 1

objective of this paper. Initially, spatial data was obtained from several official sources (e.g. 2

CTPP, NHTS, CBN, NLCD) and then manipulated to obtain spatio-temporal estimates of 3

residents, workers, students, shoppers, tourists, and stay-at-homers. The dasymetric procedure 4

developed in this study was then used to distribute this information to smaller geographic units. 5

The process can be applied to any area where the necessary input data can be obtained. 6

The modified dasymetric mapping method was applied to New Orleans in a case study using 7

ArcGIS as the platform. First, the results show population distribution varies significantly by 8

time period. Second, population was distributed in a plausible manner. Third, distinguishing 9

industrial areas from other high intensity land use avoids distributing residents, shoppers, and 10

tourists to incorrect areas. Fourth, with employee ancillary data, distribution of shopper and 11

nighttime tourists are controlled at least at the zip code level. 12

All the above information is helpful in identifying no-notice or short-notice evacuation demand. 13

Emergency managers, transportation providers, and emergency agencies such as the Red Cross 14

would benefit from such information. Further study may be needed to adjust the process to make 15

it more precise. This can be achieved by checking input data with Google Map, using more 16

ancillary data like street maps, and nighttime satellite light maps, etc. Also, more detailed 17

divisions on the temporal dimension can be done as NHTS provides data at an hourly level for 18

different trip purposes. 19

ACKNOWLEDGEMENTS 20 This research is funded by Louisiana Department of Transportation and Development (LA 21

DOTD). And we appreciate kind suggestions provided by officials to suit local needs. 22

23

24

REFERENCES 25

[1] R. D. Knabb, J. R. Rhome, and D. P. Brown, “Tropical Cyclone Report: Hurricane 26

Katrina 23-30 August 2005,” 2005. 27

[2] A. Plyer, “Facts for Features: Hurricane Katrina Impact,” The Data Center, 2013. 28

[Online]. Available: http://www.datacenterresearch.org. 29

[3] I. Eckerman, “Bhopal Gas Catastrophe 1984: Causes and Consequences,” in Reference 30

Module in Earth Systems and Environmental Sciences, Elsevier, 2013. 31

[4] FEMA, “Disaster Declarations by Year,” fema.gov, 2014. [Online]. Available: 32

http://www.fema.gov/disasters/grid/year. [Accessed: 07-Jun-2014]. 33

[5] W. and W. S. Office of Climate, “Natural Hazard Statistics,” National Weather Service, 34

2014. [Online]. Available: http://www.nws.noaa.gov/om/hazstats.shtml. [Accessed: 07-35

Jun-2014]. 36

Bian and Wilmot 16

[6] T. Abdulazim, “Using Smartphones and Sensor Technologies to Automate the Collection 1

of Travel Data,” in Transportation Research Board, 2014, vol. 175, pp. 1–16. 2

[7] A. Smith, “Smartphone Ownership 2013 | Pew Research Center’s Internet & American 3

Life Project,” 2014. [Online]. Available: 4

http://www.pewinternet.org/2013/06/05/smartphone-ownership-2013/. [Accessed: 03-5

Nov-2014]. 6

[8] S. Freire, “Modeling Daytime and Nighttime Population Distributions in Portugal Using 7

Geographic Information Systems,” University of Kansas, 2007. 8

[9] S. Freire, “Modeling of Spatiotemporal Distribution of Urban Population at High 9

Resolution – Value for Risk Assessment and Emergency Management,” in Geographic 10

Information and Cartography for Risk and Crisis Management: Towards Better Solutions, 11

M. Konecny, S. Zlatanova, and T. L. Bandrova, Eds. Berlin, Heidelberg: Springer Berlin 12

Heidelberg, 2010, pp. 53–67. 13

[10] S. K. Smith and M. House, “Temporary migration: a case study of Florida,” Popul. Res. 14

Policy Rev., vol. 26, no. 4, pp. 437–454, Jul. 2007. 15

[11] W. Kellens, T. Neutens, P. Deckers, J. Reyns, and P. Maeyer, “Coastal flood risks and 16

seasonal tourism: analysing the effects of tourism dynamics on casualty calculations,” 17

Nat. Hazards, vol. 60, no. 3, pp. 1211–1229, Aug. 2011. 18

[12] Z. Zhang, R. Sunila, and K. Virrantaus, “A spatio-temporal population model for 19

alarming, situational picture and warning system,” Remote Sens. Spat. Inf. Sci., vol. 38, 20

2010. 21

[13] T. Ahola, K. Virrantaus, J. M. Krisp, and G. J. Hunter, “A spatio-temporal population 22

model to support risk assessment and damage analysis for decision-making,” Int. J. 23

Geogr. Inf. Sci., vol. 21, no. 8, pp. 935–953, Sep. 2007. 24

[14] K. D. Bell, “Comparing Methods for Estimation of Daytime Population in Downtown 25

Indianapolis, Indiana,” Indiana University, 2011. 26

[15] J. E. Dobson, E. A. Bright, P. R. Coleman, R. C. Durfee, and B. A. Worley, “LandScan: A 27

global population database for estimating populations at risk,” Photogramm. Eng. Remote 28

Sensing, vol. 66, no. 7, pp. 849–857, 2000. 29

[16] T. N. McPherson and M. J. Brown, “Estimating daytime and nighttime population 30

distributions in US cities for emergency response activities,” in Symposium on 31

Planning,Nowcasting, and Forecasting in the Urban Zone, 84th AMS Annual Meeting, 32

2004. 33

Bian and Wilmot 17

[17] T. N. McPherson, A. Ivey, B. Michael, and S. Gerald, “Determination of the spatial and 1

temporal distribution of population for air toxics exposure assessments.pdf,” in Fifth 2

Conference on Urban Environment, 2004, p. 11. 3

[18] T. N. McPherson, J. F. Rush, H. Khalsa, A. Ivey, and M. J. Brown, “A Day-Night 4

Population Exchange Model for Better Exposure and Consequence Management 5

Assessments,” in Symposium on the Urban Environment,86th AMS Annual Meeting, 2006, 6

p. 6. 7

[19] S. Freire and C. Aubrecht, “Towards improved risk assessment: Mapping spatio-temporal 8

distribution of human exposure to earthquake hazard in the Lisbon Metropolitan Area,” in 9

Remote Sensing and Geo-Information For Environmental Emergencies, 2010. 10

[20] S. Freire and C. Aubrecht, “Assessing spatio-temporal population exposure to tsunami 11

hazard in the lisbon metropolitan area,” in the 8th International ISCRAM Conference, 12

2011. 13

[21] S. Freire and C. Aubrecht, “Integrating population dynamics into mapping human 14

exposure to seismic hazard,” Nat. Hazards Earth Syst. Sci., vol. 12, no. 11, pp. 3533–15

3543, Nov. 2012. 16

[22] R. Sleeter and N. Wood, “Estimating daytime and nighttime population density for coastal 17

communities in Oregon,” in Urban and Regional Information Systems Association Annual 18

Conference Proceedings, 2006, pp. 1–15. 19

[23] C. Aubrecht, D. Özceylan, K. Steinnocher, and S. Freire, “Multi-level geospatial modeling 20

of human exposure patterns and vulnerability indicators,” Nat. Hazards, vol. 68, no. 1, pp. 21

147–163, Sep. 2012. 22

[24] B. Bhaduri, “Population Distribution During the Day.” pp. 880–885, 2008. 23

[25] T. Kobayashi, R. M. Medina, and T. J. Cova, “Visualizing Diurnal Population Change in 24

Urban Areas for Emergency Management,” Prof. Geogr., vol. 63, no. 1, pp. 113–130, Feb. 25

2011. 26

[26] S. Holloway, J. Schumacher, and R. Redmond, “People & Place: Dasymetric Mapping 27

Using Arc/Info,” in Cartographic Design Using ArcView and Arc/Info, Missoula: 28

University of Montana, Wildlife Spatial Analysis Lab, 1997. 29

[27] J. Mennis, “Generating Surface Models of Population Using Dasymetric Mapping,” Prof. 30

Geogr., vol. 55, no. 1, pp. 31–42, 2003. 31

[28] R. Sleeter and M. Gould, “Geographic Information System Software to Remodel 32

Population Data Using Dasymetric Mapping Methods,” 2007. 33

Bian and Wilmot 18

[29] B. Bhaduri, E. A. Bright, P. R. Coleman, and M. L. Urban, “LandScan USA: a high-1

resolution geospatial and temporal modeling approach for population distribution and 2

dynamics,” GeoJournal, vol. 69, no. 1–2, pp. 103–117, Sep. 2007. 3

[30] Emergency Response Guidebook 2012. U.S. Department of Transportation, 2012. 4

5

6

7